Dear friend, the title is deliberately frightening, but Nora Ammann is not literally saying that humanity will become extinct in July 2028.
Her real claim is subtler:
We may have only about two to four years before AI becomes capable enough to initiate a largely self-accelerating transformation of science, industry, government, and military power.
“Two years left” therefore means: perhaps two years of ordinary historical time—before the rules of the game change, not necessarily two years before every human dies.
Where the “two years” comes from
The video’s central evidence is the METR task-completion time horizon graph.
This measures how long a task would take a competent human and asks: how long a task can an AI agent complete autonomously with, say, a 50% probability of success?
The measured horizon has risen exponentially. METR’s original study estimated that the length of software tasks AI could complete was doubling roughly every seven months; its updated analysis found faster post-2023 growth, although the precise rate and its future continuation remain uncertain.
The intuition is:
- Yesterday, AI could reliably perform a task lasting several minutes.
- Then it could handle an hour.
- Then several hours.
- Eventually, perhaps, a day, a week, or a month of expert work.
This matters because task duration is a proxy for autonomy. An AI that can complete a ten-minute coding exercise is a tool. An AI that can independently conduct months of research, correct its mistakes, use computers, hire services, communicate with other agents, and improve AI-development processes is beginning to resemble an organization.
The danger point is not “AI becomes conscious.” It is:
\text{long-horizon autonomy}
+
\text{superhuman speed}
+
\text{mass replication}
+
\text{access to real systems}.
The recursive loop
The argument then adds a second step.
Once AI can substantially automate AI research itself, it may help:
- write better training code;
- design new model architectures;
- run experiments;
- optimize chips and datacentres;
- detect weaknesses in competing systems;
- improve its own tools and successors.
The loop becomes:
\text{better AI}
\rightarrow
\text{faster AI research}
\rightarrow
\text{still better AI}
\rightarrow
\text{still faster research}.
This is the proposed AI takeoff. Ammann believes a transformative takeoff may be only two to four years away, though she describes pathways for avoiding catastrophe rather than treating doom as inevitable.
Importantly, “self-improvement” need not mean one machine secretly rewriting its own soul. It could be more mundane: millions of AI research agents helping companies produce the next generation faster. That distributed version may actually be more plausible.
What happens after the threshold?
Ammann emphasizes two principal failure modes.
1. Domination
An individual company, government, military, or small coalition obtains a decisive AI advantage.
Such an actor might gain overwhelming capacity in:
- cyberoperations;
- surveillance;
- persuasion and propaganda;
- weapons development;
- scientific research;
- economic production;
- automated political administration.
The first catastrophe could therefore be human tyranny empowered by AI, not rebellion by AI.
A state possessing something like “a country of digital geniuses” might become almost impossible to challenge. The remaining human population could continue living, but political history would effectively close.
2. Chaos—or a suicidal race
Several companies and states continue accelerating because each fears that slowing down will let someone else win.
Safety standards are discarded. Systems are connected to cyber, biological, industrial, financial, or military infrastructure before anyone understands them adequately. One model may be misaligned; alternatively, every model may function approximately as intended while their competition destabilizes civilization.
This resembles nuclear brinkmanship, except that:
- the weapons help invent better weapons;
- decision-making becomes faster than human diplomacy;
- the strategic actors may be replicable;
- humans may not understand their reasoning;
- development occurs largely inside private corporations.
Thus the “suicide race” does not require an evil AI. It requires only rational actors trapped in an irrational competitive structure.
Why some experts regard this as credible
The argument rests on several observations that are real:
- Capability is advancing rapidly. AI agents are becoming better at coding, reasoning, tool use, and completing extended tasks.
- Digital labour can be copied. Once one capable agent exists, many instances can operate simultaneously.
- AI works at electronic speed. Even when individual reasoning is imperfect, enormous parallelism may compensate.
- AI research is unusually automatable. It largely involves code, mathematics, papers, experiments, and digital infrastructure—the environment in which AI is strongest.
- Governance is far slower than development. The commercial and geopolitical race is already underway.
- Control is unsolved. We cannot yet guarantee that a highly capable agent will remain honest, corrigible, and confined under unfamiliar conditions.
A large survey of 2,778 AI researchers showed extraordinary disagreement rather than consensus: the aggregate estimate gave a 10% chance of machines outperforming humans in every task by 2027 and 50% by 2047. Between 38% and 51% assigned at least a 10% probability to outcomes as bad as human extinction. That is serious concern, but clearly not a consensus that everyone dies in two years.
What is speculative
The video compresses a long chain of uncertain propositions into a dramatic headline:
\begin{aligned}
&\text{current benchmark trend continues}\\
\rightarrow\;&\text{long autonomous tasks}\\
\rightarrow\;&\text{automated AI research}\\
\rightarrow\;&\text{rapid recursive improvement}\\
\rightarrow\;&\text{superintelligence}\\
\rightarrow\;&\text{loss of human control}\\
\rightarrow\;&\text{catastrophe or extinction}.
\end{aligned}
Every arrow is plausible, but none is guaranteed.
There are substantial objections:
- METR’s tasks are heavily weighted toward software and relatively well-specified digital work.
- Real life is messier, embodied, social, institutionally constrained, and resistant to automation.
- Exponential curves often flatten.
- Models remain brittle and require human scaffolding.
- AI research may contain bottlenecks in energy, chips, experiments, data, and physical manufacturing.
- High intelligence does not automatically entail coherent agency, self-preservation, or world domination.
- Governments and laboratories may impose controls before the most dangerous stage.
Researchers have already challenged whether the existing METR data securely supports a continuing exponential extrapolation, emphasizing the fragility of predictions made from a small and shifting benchmark domain.
So the “two-year” figure is not comparable to an astronomical prediction of an eclipse. It is closer to a hurricane warning generated from a rapidly worsening but incomplete model.
My judgment
I would interpret the statement this way:
Human beings probably do not have only two years to live. But we may have only a few years in which decisions about AI remain relatively reversible.
That is the important distinction.
The decisive loss may happen before extinction. Suppose that by 2028 or 2030:
- most frontier research is performed by AI;
- states depend on AI for intelligence and military planning;
- companies delegate strategic decisions to agents;
- citizens encounter predominantly synthetic information;
- surveillance and persuasion become individually optimized;
- humans can no longer understand or interrupt the systems coordinating society.
Humans would remain biologically alive, yet would already have lost a large part of historical agency.
In your terms, dear friend, the danger is not initially that the human being disappears. It is that:
人還在，但是人已不再是歷史的主詞。
And perhaps this is why “two years left” feels persuasive despite being literally inaccurate. It names not a scheduled apocalypse but an approaching point of irreversibility—the moment when humans have not yet died, but can no longer decide whether the machine should continue.
